Log-structured Merge Research Articles

Comparison and evaluation of state-of-the-art LSM merge policies

Modern NoSQL database systems use log-structured merge (LSM) storage architectures to support high write throughput. LSM architectures aggregate writes in a mutable MemTable (stored in memory), which is regularly flushed to disk, creating a new immutable file called an SSTable. Some of the SSTables are chosen to be periodically merged—replaced with a single SSTable containing their union. A mergepolicy (a.k.a. compaction policy) specifies when to do merges and which SSTables to combine. A bounded depth merge policy is one that guarantees that the number of SSTables never exceeds a given parameter k, typically in the range 3–10. Bounded depth policies are useful in applications where low read latency is crucial, but they and their underlying combinatorics are not yet well understood. This paper compares several bounded depth policies, including representative policies from industrial NoSQL databases and two new ones based on recent theoretical modeling, as well as the standard Tiered policy and Leveled policy. The results validate the proposed theoretical model and show that, compared to the existing policies, the newly proposed policies can have substantially lower write amplification with comparable read amplification.

OurRocks: Offloading Disk Scan Directly to GPU in Write-Optimized Database System

The log structured merge (LSM) tree has been widely adopted by database systems owing to its superior write performance. However, LSM-tree based databases face vulnerabilities when processing analytical queries due to the read amplification caused by its architecture and the limited use of storage devices with high bandwidth. To flexibly handle transactional and analytical workloads, we proposed and implemented OurRocks taking full advantage of NVMe SSD and GPU devices, which improves scan performance. Although the NVMe SSD serves multi GB/s I/O rates, it is necessary to solve the data transfer overhead which limits the benefits of the GPU processing. The primary idea is to offload the scan operation to the GPU with filtering predicate pushdown and resolve the bottleneck from the data transfer between devices with direct memory access (DMA). OurRocks benefits from all the features of write-optimized database systems, in addition to accelerating the analytic queries using the aforementioned idea. Experimental results indicate that OurRocks effectively leverages resources of the NVMe SSD and GPU and significantly improves the execution of queries in the YCSB and TPC-H benchmarks, compared to the conventional write-optimized database. Our research demonstrates that the proposed approach can speed up the handling of the data-intensive workloads.

Cascaded Write Amplification of LSM-tree-based Key-Value Stores underlying Solid-State Disks

Log-structured merge tree (i.e., LSM-tree)-based key–value stores (i.e., KV stores) are widely used in big-data applications and provide high performance. NAND Flash-based Solid-state disks (i.e., SSDs) have become a popular storage device alternative to hard disk drives (i.e., HDDs) because of their high performance and low power consumption. LSM-tree KV stores with SSDs are deployed in large-scale storage systems, which aims to achieve high performance in the cloud. Write amplification in LSM-tree KV stores and NAND Flash memory in SSDs are defined as WA1 and WA2 in this paper. The former, which is attributed to compaction operations in LSM-tree-based KV stores, is a burden on I/O bandwidth between the host and the device. The latter, which results from out-place updates in NAND Flash memory, blocks user I/O requests between the host and NAND Flash memory, thereby degrading the SSD performance. Write amplification impairs the overall system performance. In this study, we explored the two-level cascaded write amplification in LSM-tree KV stores with SSDs. The cascaded write amplification is represented as WA. Our primary goal is to comprehensively study two-level cascaded write amplification on the host-side LSM-tree KV stores and the device-side SSDs. We quantitatively analyze the impact of two-level write amplification on overall performance. The cascaded write amplification is 16.44 (WA1 is 16.55; WA2 is 0.99) and 35.51 (WA1 is 16.6; WA2 is 2.14) for SSD-I and SSD-S with LevelDB’s default setting under DB_bench. The larger cascaded write amplification of KV stores has a bad impact on SSD performance and lifetime. The throughput of SSD-S and SSD-I under an 80%-write workload is approximately 0.28x and 0.31x of that under a 100%-write workload. Therefore, it is important to design a novel approach to balance the cost of an SSD lifetime caused by cascaded write amplification and its high performance under the read-write-mixed workloads. We attempt to reveal details of cascaded write amplification and hope that this study is useful for developers of LSM-tree-based KV stores and SSD software stacks.

An LSM-based tuple compaction framework for Apache AsterixDB

Document database systems store self-describing semi-structured records, such as JSON, "as-is" without requiring the users to pre-define a schema. This provides users with the flexibility to change the structure of incoming records without worrying about taking the system offline or hindering the performance of currently running queries. However, the flexibility of such systems does not free. The large amount of redundancy in the records can introduce an unnecessary storage overhead and impact query performance. Our focus in this paper is to address the storage overhead issue by introducing a tuple compactor framework that infers and extracts the schema from self-describing semi-structured records during the data ingestion. As many prominent document stores, such as MongoDB and Couchbase, adopt Log Structured Merge (LSM) trees in their storage engines, our framework exploits LSM lifecycle events to piggyback the schema inference and extraction operations. We have implemented and empirically evaluated our approach to measure its impact on storage, data ingestion, and query performance in the context of Apache AsterixDB.

Integrating LSM Trees With Multichip Flash Translation Layer for Write-Efficient KVSSDs

Log-structured-merge (LSM) trees are a highly write-optimized data structure for lightweight, high-performance key-value (KV) stores. Furthermore, solid-state drives (SSDs) are a crucial component for I/O acceleration. Conventional LSM-over-SSD designs involve multiple software layers, including the LSM tree, host file system, and flash translation layer (FTL), which introduce cascading write amplifications. To manage the write amplifications from different layers, we propose KVSSDs, which are a close integration of LSM trees and the FTL. KVSSDs exploit the FTL mapping mechanism to implement copy-free compaction of LSM trees, and they enables direct data allocation in flash memory for efficient garbage collection. Our design also uses a fine-grained, dynamic striping policy to fully exploit the rich internal parallelism of multichip SSDs. The experimental results indicated that our LSM-SSD integrated design reduced the write amplification by 86% and improved the throughput by 383% compared with a conventional LSM-over-SSD design.

CaseDB: Lightweight Key-Value Store for Edge Computing Environment

Key-value stores based on a log-structured merge (LSM) tree have emerged in big data systems because of their scalability and reliability. An LSM-tree offers a multilevel data structure with a simple interface. However, it performs file rewrites at the disk level, which causes write amplification. This study is concerned with this problem in relation to an embedded board environment, which can be used in edge computing. Addressing the major problems associated with an LSM-tree, we propose a new key-value store named CaseDB, which aggressively separates keys and bloom filters on the non-volatile memory express (NVMe) drive and stores the values on the SSD. Our solution reduces the I/O cost and enhances the overall performance in a cost-efficient manner. CaseDB employs a memory component, CBuffer, to avoid small write operations, and a delayed value compaction technique that guarantees the sorted order for both keys and values. CaseDB also utilizes deduction-based data deduplication to prevent space amplification in the values layer. The experiments show that CaseDB outperforms LevelDB and WiscKey 5.7 and 1.8 times, respectively, with respect to data writes, and additionally improves the read performance by 1.5 times. CaseDB also avoids the space amplification of WiscKey.

Adaptive Lower-level Driven Compaction to Optimize LSM-Tree Key-Value Stores

Log-structured merge (LSM) tree key-value (KV) stores have been widely deployed in many NoSQL and SQL systems, serving online big data applications such as social networking, graph processing, machine learning, etc. The batch processing of sorted data merging (i.e., compaction) in LSM-tree key-value stores improves the write efficiency, and some lazy compaction methods have been proposed to accumulate more data within a batch. However, these batched writing methods lead to significant tail latency, which is unacceptable for online processing. Aiming to optimize both latency and throughput, we propose a novel Lower-level Driven Compaction (LDC) method which breaks the limitations of the traditional upper-level driven compaction manner and triggers practical compaction actions bottom-up, with the benefits of both decreasing the compaction granularity for smaller latency and reducing write amplification for higher throughput. Furthermore, we extend LDC to Adaptive LDC (ALDC) by adding an adaptive policy to adjust the key compaction threshold to fit the changes of workloads' features. The experimental results indicate that ALDC reduces the tail latency significantly and meanwhile achieves a much higher and stable throughput compared with existing approaches.

Partition pruning for range query on distributed log-structured merge-tree

Log-structured merge tree (LSM-tree) is adopted by many distributed storage systems. It contains a Memtable and a number of SSTables. The Memtable is an in-memory structure and the SSTable is a disk-based structure. Data records are horizontally partitioned over the primary key and stored in different SSTables. Data writes on records are first served by the Memtable and then compacted to SSTables periodically. Although this design optimizes data writes by avoiding random disk writes, it is unfriendly to read request since the results should be retrieved and merged from both Memtable and SSTables. In particular, when the Memtable and SSTables are distributed on different nodes, it incurs expensive costs to serve range queries. A range query on non-primary key columns has to scan all partitions, which generates many network and I/O expenses. In this paper, we propose a partition pruning strategy to save cost for range queries. A statistics cache is designed to determine whether a partition contains the desired data or not, which enables read requests to avoid scanning useless partitions. As records can be updated in Memtable freely, to prevent incorrect filtering, a version-based cache synchronization strategy is proposed to ensure the queries to obtain the latest data state. We implement the proposed method in an open source distributed database and conduct comprehensive experiments. Experimental results reveal that the performance of range queries increased 30% ~ 40% with our partition pruning technique.

TeksDB

Key-value stores (KVS) are now an integral part of modern dataintensive systems. thanks to its simplicity, scalability, and efficiency over traditional database systems. Databases such as MySQL employ KVS (in this case, RocksDB as their backend storage instead of conventional database engines because KVS is less affected by data fragmentation. OLAP and big data analytics applications often use KVS as the backend store to make use of the efficient range_query operations of LSM(Log Structured Merge Tree)-based KVS. Furthermore, the increasing use of KVS as backend stores in distributed systems diversifies the workload for each local KVS, making it necessary to design a highly-concurrent data store that simultaneously processes mixed operations. The underlying technology trends also put immense pressure on the design of KVS. Nowadays, many core servers with hundred GBs of memory are common in production environments, and this enhanced density of hardware technologies (both in terms of the number of cores and memory capacity) leads to more data items residing in memory. This phenomenon places greater emphasis on the efficiency of the in-memory data structure in KVS, as more data accesses are serviced from its memory component, which is often called the memtable. The current memtable structure of KVS, however, generally assumes a limited memory capacity that only serves as a temporary write buffer, and the use of single-threaded writes or unscalable data structures for the memtable fails to fulfill this external demand.

SEALDB: An Efficient LSM-tree Based KV Store on SMR Drives with Sets and Dynamic Bands

Key-value (KV) stores play an increasingly critical role in supporting diverse large-scale applications in modern data centers hosting terabytes of KV items which even might reside on a single server due to virtualization purposes. The combination of the ever-growing volume of KV items and storage/application consolidation is driving a trend of high storage density for KV stores. Shingled Magnetic Recording (SMR) represents a promising technology for increasing disk capacity, which however comes with the increased complexity of handling random writes. To take the best advantages of SMR drives, applications are expected to work in an SMR-friendly way. In this work, we present SEALDB, a Log-Structured Merge tree (LSM-tree) based key-value store that is specifically optimized for SMR drives via avoiding random writes and the corresponding write amplification on SMR drives. First, for LSM-trees, SEALDB collects and groups participating data of each compaction into sets. Using a set as the basic unit for compactions, SEALDB improves compaction efficiency by reducing random I/Os. Second, SEALDB creates variable sized bands on original HM-SMR drives, named dynamic bands. Dynamic bands store sets in an SMR-friendly way to eliminate the auxiliary write amplification from SMR drives. Third, SEALDB employs two light-weight garbage collection (GC) policies to further improve the space efficiency. We demonstrate the advantages of SEALDB via extensive experiments with various workloads. Overall, SEALDB delivers impressive performance compared with LevelDB, e.g., $3.42\times$3.42×/$2.65\times$2.65× faster for random writes (without or with GCs), and $3.96\times$3.96× faster for sequential reads.

Enabling Efficient Updates in KV Storage via Hashing

Persistent key-value (KV) stores mostly build on the Log-Structured Merge (LSM) tree for high write performance, yet the LSM-tree suffers from the inherently high I/O amplification. KV separation mitigates I/O amplification by storing only keys in the LSM-tree and values in separate storage. However, the current KV separation design remains inefficient under update-intensive workloads due to its high garbage collection (GC) overhead in value storage. We propose HashKV, which aims for high update performance atop KV separation under update-intensive workloads. HashKV uses hash-based data grouping , which deterministically maps values to storage space to make both updates and GC efficient. We further relax the restriction of such deterministic mappings via simple but useful design extensions. We extensively evaluate various design aspects of HashKV. We show that HashKV achieves 4.6× update throughput and 53.4% less write traffic compared to the current KV separation design. In addition, we demonstrate that we can integrate the design of HashKV with state-of-the-art KV stores and improve their respective performance.

Fault-tolerant precise data access on distributed log-structured merge-tree

Log-structured merge tree has been adopted by many distributed storage systems. It decomposes a large database into multiple parts: an in-writing part and several read-only ones. Records are firstly written into a memory-optimized structure and then compacted into in-disk structures periodically. It achieves high write throughput. However, it brings side effect that read requests have to go through multiple structures to find the required record. In a distributed database system, different parts of the LSM-tree are stored in distributed fashion. To this end, a server in the query layer has to issues multiple network communications to pull data items from the underlying storage layer. Coming to its rescue, this work proposes a precise data access strategy which includes: an efficient structure with low maintaining overhead designed to test whether a record exists in the in-writing part of the LSM-tree; a lease-based synchronization strategy proposed to maintain consistent copies of the structure on remote query servers.We further prove the technique is capable of working robustly when the LSM-Tree is re-organizing multiple structures in the backend. It is also fault-tolerant, which is able to recover the structures used in data access after node failures happen. Experiments using the YCSB benchmark show that the solution has 6x throughput improvement over existing methods.

Collaborative Compaction Optimization System using Near-Data Processing for LSM-tree-based Key-Value Stores

Log-structured merge tree (LSM-tree) based key–value stores are widely employed in large-scale storage systems. In compaction, high-level sorted string table files (i.e., SSTables) are merged with low-level overlapping key ranges and sorted for data queries. However, the compaction process incurs write amplification, which degrades system performance particularly under update-intensive workloads. Current optimizations mostly focus on reducing the overload of compaction in the host but rarely exploit parallelisms between the host and the device to fully utilize computing resources in the entire system for optimizing compaction performance. In this study, we propose Co-KV, a Collaborative Key-Value store, to improve compaction performance in LSM-tree-based key–value stores. Co-KV is based on a near-data processing (i.e., NDP) model-enabled storage device. Co-KV exhibits the following advantages: (1) it reduces write amplification and host-side CPU costs using a compaction offloading scheduling between a host computer and an NDP-enabled storage device; (2) it relieves the overload associated with data transfer between the host and the storage device; and (3) it improves the compaction of LSM-tree based key–value stores under update-intensive workloads.We employed an Open Ethernet Driver (OED), which is a real-world NDP platform as the testbed for our experiments. Extensive db_bench evaluations demonstrate that Co-KV achieves overall throughput improvements of approximately 1.75x, CPU cost reductions of approximately 68.1%, and write amplification reductions by up to 50.0% over the state-of-the-art LevelDB. Under YCSB workloads, Co-KV increases throughput by 1.7x ∼ 1.9x while decreasing the write amplification and average latency by up to 50.0% and 46.3%, respectively.

범위 기반 로그 구조 병합 트리를 활용한 키-값 저장소의 최적화 방법

범위 기반 로그 구조 병합 트리를 활용한 키-값 저장소의 최적화 방법

Efficient data ingestion and query processing for LSM-based storage systems

In recent years, the Log Structured Merge (LSM) tree has been widely adopted by NoSQL and NewSQL systems for its superior write performance. Despite its popularity, however, most existing work has focused on LSM-based key-value stores with only a single LSM-tree; auxiliary structures, which are critical for supporting ad-hoc queries, have received much less attention. In this paper, we focus on efficient data ingestion and query processing for general-purpose LSM-based storage systems. We first propose and evaluate a series of optimizations for efficient batched point lookups, significantly improving the range of applicability of LSM-based secondary indexes. We then present several new and efficient maintenance strategies for LSM-based storage systems. Finally, we have implemented and experimentally evaluated the proposed techniques in the context of the Apache AsterixDB system, and we present the results here.

SILK+ Preventing Latency Spikes in Log-Structured Merge Key-Value Stores Running Heterogeneous Workloads

Log-Structured Merge Key-Value stores (LSM KVs) are designed to offer good write performance, by capturing client writes in memory, and only later flushing them to storage. Writes are later compacted into a tree-like data structure on disk to improve read performance and to reduce storage space use. It has been widely documented that compactions severely hamper throughput. Various optimizations have successfully dealt with this problem. These techniques include, among others, rate-limiting flushes and compactions, selecting among compactions for maximum effect, and limiting compactions to the highest level by so-called fragmented LSMs. In this article, we focus on latencies rather than throughput. We first document the fact that LSM KVs exhibit high tail latencies. The techniques that have been proposed for optimizing throughput do not address this issue, and, in fact, in some cases, exacerbate it. The root cause of these high tail latencies is interference between client writes, flushes, and compactions. Another major cause for tail latency is the heterogeneous nature of the workloads in terms of operation mix and item sizes whereby a few more computationally heavy requests slow down the vast majority of smaller requests. We introduce the notion of an Input/Output (I/O) bandwidth scheduler for an LSM-based KV store to reduce tail latency caused by interference of flushing and compactions and by workload heterogeneity. We explore three techniques as part of this I/O scheduler: (1) opportunistically allocating more bandwidth to internal operations during periods of low load, (2) prioritizing flushes and compactions at the lower levels of the tree, and (3) separating client requests by size and by data access path. SILK+ is a new open-source LSM KV that incorporates this notion of an I/O scheduler.

PLI $$^+$$ + : efficient clustering of cloud databases

Commercial cloud database services increase availability of data and provide reliable access to data. Routine database maintenance tasks such as clustering, however, increase the costs of hosting data on commercial cloud instances. Clustering causes an I/O burst; clustering in one-shot depletes I/O credit accumulated by an instance and increases the cost of hosting data. An unclustered database decreases query performance by scanning large amounts of data, gradually depleting I/O credits. In this paper, we introduce Physical Location Index Plus ( $${PLI}^{\small {{+}}}$$ ), an indexing method for databases hosted on commercial cloud. $${PLI}^{\small {{+}}}$$ relies on internal knowledge of data layout, building a physical location index, which maps a range of physical co-locations with a range of attribute values to create approximately sorted buckets. As new data is inserted, writes are partitioned in memory based on incoming data distribution. The data is written to physical locations on disk in block-based partitions to favor large granularity I/O. Incoming SQL queries on indexed attribute values are rewritten in terms of the physical location ranges. As a result, $${PLI}^{\small {{+}}}$$ does not decrease query performance on an unclustered cloud database instance, DBAs may choose to cluster the instance when they have sufficiently large I/O credit available for clustering thus delaying the need for clustering. We evaluate query performance over $${PLI}^{\small {{+}}}$$ by comparing it with clustered, unclustered (secondary) indexes, and log-structured merge trees on real datasets. Experiments show that $${PLI}^{\small {{+}}}$$ significantly delays clustering, and yet does not degrade query performance—thus achieving higher level of sortedness than unclustered indexes and log-structured merge trees. We also evaluate the quality of clustering by introducing a measure of interval sortedness, and the size of index.

High-throughput publish/subscribe on top of LSM-based storage

State-of-the-art publish/subscribe systems are efficient when the subscriptions are relatively static—for instance, the set of followers in Twitter—or can fit in memory. However, now-a-days, many big data and IoT based applications follow a highly dynamic query paradigm, where both continuous queries and data entries are in the millions and can arrive and expire rapidly. In this paper we propose and compare several publish/subscribe storage architectures, based on the popular NoSQL log-structured merge tree (LSM) storage paradigm, to support high-throughput and highly dynamic publish/subscribe systems. Our framework naturally supports subscriptions on both historic and future streaming data, and generates instant notifications. We also extend our framework to efficiently support self-joining subscriptions, where streaming pub/sub records join with past pub/sub entries. Further, we show how hierarchical attributes, such as concept ontologies, can be efficiently supported; for example, a publication’s topic is “politics” whereas a subscription’s topic is “US politics.” We implemented and experimentally evaluated our methods on the popular LSM-based LevelDB system, using real datasets, for simple match and self-joining subscriptions on both flat and hierarchical attributes. Our results show that our approaches achieve significantly higher throughput compared to state-of-the-art baselines.

Accordion

Log-structured merge (LSM) stores have emerged as the technology of choice for building scalable write-intensive key-value storage systems. An LSM store replaces random I/O with sequential I/O by accumulating large batches of writes in a memory store prior to flushing them to log-structured disk storage; the latter is continuously re-organized in the background through a compaction process for efficiency of reads. Though inherent to the LSM design, frequent compactions are a major pain point because they slow down data store operations, primarily writes, and also increase disk wear. Another performance bottleneck in today's state-of-the-art LSM stores, in particular ones that use managed languages like Java, is the fragmented memory layout of their dynamic memory store. In this paper we show that these pain points may be mitigated via better organization of the memory store. We present Accordion - an algorithm that addresses these problems by re-applying the LSM design principles to memory management. Accordion is implemented in the production code of Apache HBase, where it was extensively evaluated. We demonstrate Accordion's double-digit performance gains versus the baseline HBase implementation and discuss some unexpected lessons learned in the process.

Algorithms Behind Modern Storage Systems

This article takes a closer look at two storage system design approaches used in a majority of modern databases (read-optimized B-trees and write-optimized LSM (log-structured merge)-trees) and describes their use cases and tradeoffs.